A Deep Learning Approach for Fishing Vessel Classification from VMS Trajectories Using Recurrent Neural Networks

Abstract

Satellite-based vessel monitoring systems (VMS) have been widely deployed on fishing vessels for monitoring and surveillance. In this study, we aim to enhance the classification of fishing ship trajectory from the VMS data. We propose a recurrent neural network (RNN)-based approach for discrimination of fishing vessel types from ship trajectories. Our proposed method first eliminates data points that are meaningless by identifying groups of data points describing ship movements using a density-based clustering strategy. We then generate local trajectories and compute a feature vector for each identified group as input for RNN. Finally, we train RNN models to learn high-level representation of ship trajectory for the task of classification. Experiments conducted on real-world VMS records among three fishing ship types: trawl, purse seine, and falling net demonstrate the effective use of RNNs and bidirectional GRU performs the best performance with 89.74% accuracy.